Meta has introduced two innovative large models, namely Muse Wonder and Muse Wonder Full. According to the company, these models have been overhauled with a reconstructed underlying technology stack, and they have showcased exceptional performance in a variety of self-constructed and third-party tests. These tests span areas such as agent capabilities, proactive memory, long-term software engineering, and legal vertical domains, outperforming leading-edge models like GPT-6 Astra and Claude Opus 5.5. Moreover, they incorporate core functionalities such as proactive memory and focus. However, following the opening of API access, numerous developers discovered through hands-on testing that the models' performance in 3D game-related scenarios significantly trailed behind their competitors. This discrepancy has sparked suspicions of targeted benchmark manipulation. In response, Meta clarified that they did not specifically train the models for 3D game-related content and conceded that, compared to competitors, the models underperform in certain specific scenarios. Additionally, the most powerful version of the models is not yet available for API calls. Furthermore, the API pricing for these models is relatively high, and they are no longer open-source. A Meta executive disclosed that over $100 billion has been invested in infrastructure to develop these models. They also put forward industry perspectives, including the necessity for substantial upfront investment in cutting-edge AI, viewing models as products, and the anticipated shift towards an AI agent B2B era. At present, the true overall capabilities of these models still necessitate more practical testing for thorough verification.